Four losses that no tool undoes

Knowing this is a professional skill

The most expensive mistake in this trade is spending four hours on a file that could not be fixed, and telling the client on Friday. Recognising an unrecoverable file in the first five minutes is worth more than any technique in this course.

There are four main losses, and each has a mechanism that explains why it is permanent.

1. Motion blur and camera shake

A blurred photograph is the sharp image convolved with a blur kernel: the path the camera or subject travelled during the exposure. Undoing it is deconvolution.

If you know the kernel exactly, deconvolution works mathematically. Astronomers and microscopists do it routinely because they can measure their optical system’s point spread function precisely.

For an ordinary photograph you do not know the kernel, so the problem is blind deconvolution: estimating both the kernel and the sharp image from the blurred result alone. That is ill-posed (many combinations of kernel and image produce the same blur) and in the presence of noise the estimate amplifies the noise enormously.

What this means in practice:

• Simple, short, linear motion blur can be partially undone. Photoshop’s Shake Reduction and GIMP’s G’MIC deconvolution filters do recover some sharpness on mild cases.
• Anything more (long exposure, rotational shake, a subject moving while the camera moved, multiple objects at different speeds) produces ringing artefacts worse than the blur.
• Model-based deblurring produces a sharp-looking image by generating plausible detail. It is not recovery; it is the same mechanism as AI upscaling, with the same consequences for anything evidential.

2. Out of focus

Same mathematics, different kernel (a disc rather than a streak) and worse, because defocus blur discards information more thoroughly. Fine detail below the blur’s scale has been averaged to a uniform value, and an average does not retain what was averaged.

An additional problem: focus varies with depth. A photograph focused on the background has different blur on the foreground, so there is no single kernel for the frame.

Sharpening does not help, because sharpening exaggerates edges that exist and a defocused image’s edges are gone. Applying heavy sharpening to a soft photograph gives you a soft photograph with halos.

The practical answer is always the same: if the eyes are not sharp, the frame is not usable for anything that will be seen large. It may still be fine at 400 pixels on a web page, and that is a legitimate use.

3. Clipped highlights and crushed blacks

When all three channels reach 255, every pixel in that region holds the same value. A blown sky is a uniform patch of 255 with no variation: there is nothing to recover because nothing distinguishes one pixel from another.

Raw gives you the partial case: where one or two channels clipped and one did not, the missing ones can be reconstructed. That is the genuine one-to-two-stop headroom from the raw lesson. Where all three are at maximum, no format helps.

What can be done is replacement: putting a sky from another frame in, or painting a gradient. That is compositing, it should look deliberate, and it should be disclosed if the picture is documentary.

Crushed blacks are the same at the other end, and they are usually less damaging because the eye expects shadows to contain little.

4. Compression and resolution

A heavily compressed JPEG has had its transform coefficients rounded to zero. A small image never sampled the detail. Both are absences, and both attract tools that claim to restore them.

Deblocking filters smooth the artefacts, which trades one loss for another. Model-based restoration generates plausible detail from the low-quality evidence. As established in the first module, that is invention that looks like recovery, and the failure modes are specific: letterforms come back as the wrong letters, faces come back as slightly different people, repeated textures come back as convincing patterns that are not the pattern that was there.

For a family photograph destined for a frame, plenty of people are happy with the result and should be told what it is. For a document, a number plate, a medical image, a legal exhibit or a news photograph, it is not usable, because what comes out is a guess wearing the appearance of a fact.

The five-minute triage

Before accepting a repair job, check in this order:

1. Open at 100 per cent and look at the most important detail: usually the eyes, the label, the face. Sharp or not sharp? This is the fastest single test.
2. Histogram, three-channel. Are the highlights piled at 255 across all three? Are the shadows at 0?
3. Zoom to 200 per cent on a flat area. JPEG blocks? Heavy noise? Both cap what the file can become.
4. Pixel dimensions against the intended output. Divide the pixel width by the printed inches. Under 120 and you have a conversation to have.
5. Ask whether an original exists. Most “unusable” files are WhatsApp copies of perfectly good originals, and one message solves it.

That is five minutes and it converts a bad week into a clear answer on day one.

How to say it

“This can’t be fixed” is rarely the most useful sentence. Better:

“The eyes are out of focus, so it won’t hold up printed larger than about 10 by 15 centimetres. I can correct the colour and the damage, and it will look good at that size. If you need it bigger, is there another frame from the same day?”

That names the limit, states what is possible, gives a size, and asks the question that most often solves the problem. It also protects you, because the expectation is now written down.

The honest general statement to carry out of this module: editing redistributes and repairs information; it does not create it. Everything that appears to break that rule is a model inventing plausible content, and the only question worth asking about it is whether the picture is allowed to contain invention. That question is answered by what the picture is for, not by the tool.